Combining neighborhood separable subspaces for classification via sparsity regularized optimization

摘要: The neighborhood rough set theory has been successfully applied to various classification tasks. key concept of this is find a sufficient and necessary separable subspace for building compact model. Given learning task, there usually exist numerous subspaces that maintain the discriminative ability original space with respect given granularity. These contain complementary information classification. However, it challenging task compute these efficiently. In paper, we develop fast attribute reduction algorithm based on sample pair selection all reducts. Nevertheless, cannot deal large-scale data. Then propose randomized dependency. can part reducts very efficient. A framework joint representation proposed fully exploit in different subspaces. addition, weight matrix learned combine residuals via group sparsity regularization. performances algorithms are compared, influence granularity discussed. Finally, technique compared other ensemble algorithms. Experimental results show superior state-of-the-art classifiers.